The infrared image based non-contact monitoring of respiratory waveform through deep Kalman filter

Respiratory waveform is one of the most important physiological signals containing essential pathophysiological information. The classical monitoring of respiratory waveform is based on the flow meter with contacted inputs. A non-contact respiratory waveform monitoring method is needed to bring a better patient experience, allow more application scenarios and provide additional measurements to gain an in-depth understanding of the respiration system. In this paper, we proposed a novel infrared image-based non-contact monitoring method which successfully obtains the detailed preserved respiratory waveform for the first time. The obtained infrared image is modelled as temperature distribution over a spatial field instead of a simple grey image, which is decided mainly by the flow speed. And an efficient analytical model guided mapping function from raw high-dimensional observations into temporal flow sequences is developed to replace the simple average over the region of interests. As a result, the manual-involved measurement noises can be significantly suppressed. To further mitigate the residual noises, a deep Kalman filter is designed to make use of the self-evolution model of the respiration system. The experimental results have validated the accuracy of the proposed method.